High Unemployment Yet Few Small Firms: The Roleof Centralized Bargaining in South Africa

Jeremy R. Magruder

November 11, 2011

Abstract

South Africa has very high unemployment, yet few adults work informally in smallrms. This paper tests whether centralized bargaining, by which unionized largerms extend arbitration agreements to non-unionized smaller rms, contributes to thisproblem. While local labor market characteristics inuence the location of theseagreements, their coverage is spatially discontinuous, allowing identication by spatialregression discontinuity. Centralized bargaining agreements are found to decrease em-ployment in an industry by 8-13%, with losses concentrated among small rms. Theseeects are not explained by resettlement to uncovered areas, and are robust to a widevariety of controls for unobserved heterogeneity.

University of California, Berkeley. I thank Guojun He and Charles Seguin for excellent research as-sistance and John Bellows for alerting me to some of the data used in this study. I also thank MichaelAnderson, Lori Beaman, Tim Conley, Ann Harrison, Elisabeth Sadoulet, Paul Schultz, Chris Udry andseminar audiences at Berkeley, USC, UCSD, the World Bank, the University of Stellenbosch, the Universityof Cape Town, The University of the Witwatersrand, the University of Pretoria, Yale, and Northwesternfor many helpful comments, and also Economic Research Southern Africa for organizing a workshop andfacilitating seminars in South Africa. All remaining mistakes are naturally my own.

1

1 Introduction

Wage-setting institutions, including collective bargaining and labor legislation, are politically

charged issues for countries at all stages of development. Concern over the rst-order

theoretical implications of these labor market distortions that mandating improved working

conditions should induce lower employment has been somewhat attenuated by the vast

empirical literature which has examined these issues. The large majority of these studies

have taken place in OECD countries, and are characterized by a nding of small employment

eects (Blau and Kahn 1999; for a survey of evidence from developing countries see Freeman

2009). It seems likely that the labor market response to wage-setting institutions changes as

countries develop. After all, for wage-employment tradeos to be economically important,

two structural features must hold: rst, there must be substantial labor supply at low wages

(below the cut of the proposed standards), and second, the government must have the

capacity to actually enforce these regulations. We might expect labor markets in middle

income countries to exhibit these two features, and particularly so in countries with high

rates of inequality. There, a large fraction of the labor force experiences living standards

similar to those in much poorer countries and may be willing to work at very low wages.

At the same time, the governments tax base and enforcement capacity may have grown

stronger, attaching credibility to labor law1.

In South Africa, a middle income country with very high inequality, the labor market

appears heavily distorted. Similar to many other low and middle income countries, for-

mal sector work is dominated by large scale employers, where it is highly regulated, highly

remunerated, and scarce. However, very dierently than many peer countries, there are

few small rm jobs of any sort, informal or otherwise2. The outcome is an astronomical

1Indeed, previous studies of minimum wages in Latin America have provided evidence that minimumwage increases yield increases in wages in both the formal and informal sectors (Freeman 2009), suggestingthat legislated labor standards can impact the labor market across the labor supply distribution in middleincome countries.

2For an extended discussion of small rms in developing countries, see Liedholm and Mead (1987, 1999).

2

unemployment rate, where only 56% of prime-aged men and 40% of prime-aged women are

actually working. Understanding why there is so little employment in small rms is a ques-

tion of fundamental policy importance to South Africa. This is particularly true if labor

standards contribute to the problem, as then policy prescriptions are immediate. In that

case, conclusions may also have direct analogues when similar policies are implemented in

other contexts.

While there are many labor regulations in South Africa, one particular form of labor

standards which has been implicated in the lack of small rms is the bargaining council

system3. Similar to centralized bargaining structures across Western Europe (Nickell 1997)

and in Argentina and Brazil (Carneiro 1997; Cardoso and Gindin 2009), employer organi-

zations and unions may opt to participate in bargaining councils, which extend arbitration

agreements beyond the rms and unions which make them to all workers in an industry in a

given political demarcation, regardless of rm size or participation in the arbitration process.

If large rms and unions agree to high standards with the goal of reducing competition from

small rms, then this could limit the viability of small rm enterprise, restricting the options

available to the unemployed. However, despite the importance of centralized bargaining

in a variety of countries, there have been few empirical studies which have provided strong

causal evidence of the employment or industrial structure eects of centralized bargaining

agreements in any context, leaving us with scant evidence to inform policy.

Understandably, then, these bargaining councils are at the center of a vigorous policy

debate in South Africa, with small rms arguing that the labor standards impose unfair costs,

while large rms assert that these labor standards are not punitive and union alliances argue

that they are necessary for worker protection. However, clean identication of the eects

of these agreements on employment and industrial structure which could inform policy has

been elusive thus far. There are several challenges to identication. First, other potential

3Presumably, informal rms could not be compelled to comply with bargaining council agreements andthus the presence of this or any other labor law seems unlikely to be the primary reason for the absence ofinformal small rms.

3

motivations for South Africas unemployment without entrepreneurship abound, many of

which are surely empirically relevant. For example, high wages and an extensive social

safety net may increase the demand for leisure or render long periods of unemployed search

more palatable. Entrepreneurial opportunities may be limited by low skill levels, liquidity

constraints, and high crime, while high capital stocks may make large rms competitive with

relatively low labor inputs. Moreover, the legacy of Apartheid looms large; the majority

black population was prohibited from entrepreneurship during Apartheid. While these laws

are no longer in place, their eects on skillsets and culture may have a lingering impact.

Identifying the employment eects of labor regulations, therefore, requires a careful analysis

which would hold these conditions constant. Further complicating analysis, these agreements

are outcomes of a complex bargaining process between unions and rms with unclear and

likely anti-competitive motives. Since centralized bargaining is not mandated, the rms

which choose to pursue centralized bargaining may be those who work in local labor markets

where centralized bargaining would represent a particularly large competitive advantage.

This paper estimates the eects of bargaining councils using several methods, taking

advantage of the tremendous amount of variation in the data: agreements vary with space,

across industries, and over time. As a benchmark, a dierence-in-dierences estimates sig-

nicant negative eects of bargaining councils on employment and small rm employment.

As it is possible that trends in local labor market characteristics are related to secular trends

in bargaining council status, confounding that estimator, the paper then takes advantage of

the fact that local labor markets should be spatially continuous within South Africa, while

these agreements are enforced in a spatially discontinuous way. This suggests a spatial

regression discontinuity estimation. Using distance to policy regime borders as the running

variable, the spatial regression discontinuity reveals even larger negative employment eects

which are both visibly and statistically signicant. Finally, this paper adopts the spatial

xed eects proposed in Conley and Udry (2008) and Goldstein and Udry (2008) as its pre-

ferred estimator, arguing that these spatial xed eects have several advantages over a more

4

traditional regression discontinuity in a spatial context. These spatial xed eects estimates

report consistent and large eects: industries which have an agreement in a particular mag-

isterial district in a given year have about 8-13% lower employment and 10-21% higher wages

than the same industry in uncovered neighboring magisterial districts. Firm sizes are also

impacted, with 7-16% fewer employees in small rms and 7-15% fewer entrepreneurs, while

there are smaller and insignicant eects on large rms and single employee rms. Utiliz-

ing magisterial district-year and magisterial district-industry xed eects, I show that these

spatial discontinuities are similar in magnitude and precision whether only inter-industry

variation (within a magisterial district-year) or intertemporal variation (within a magisterial

district-industry) is utilized, and that estimates increase in magnitude and precision if we

consider only small magisterial districts who should be unable to endogenously inuence

these agreements. I further illustrate that, while rms do move across borders in order to

avoid these agreements, this border-jumping does not drive the employment eects measured

here, so that these reductions in employment represent a net loss for the economy. The ten

percent employment eect on covered industries is large relative to those which have been

identied in other labor regulation studies, and accounts for about a percentage point of

unemployment. These bargaining councils thus have an eect which is economically signif-

icant and should be of interest to policy makers both in South Africa and in other middle

income countries. However, they cannot explain the majority of unemployment in South

Africa, and therefore bargaining councils are best understood as exacerbating an existing

and severe problem.

2 South Africas Missing Small Firms

Unemployment in South Africa is extremely high, particularly among non-whites. The

rst two columns of Table 1 report data from the 2003 Labour Force Survey (described

below), which indicates that only about 56% of 20-60 year old men and 40% of 20-60 year

5

old women are actually working (the corresponding employment rates are 50% and 35% if

we restrict our attention to the majority black population). These numbers correspond to a

34% unemployment rate in this prime-aged population (where unemployment is dened as

wanting work)4. A large number of potential reasons for this unemployment exist, and the

unemployment numbers and potential contributors for them are surveyed more extensively in

a series of papers by Kingdon and Knight (2004, 2006, 2008) as well as Banerjee et al (2008).

Wages are high, due to high capital/labor ratios, a strong union presence, and extensive

governmental labor market regulation in addition to the industrial bargaining agreements

which are the focus of this paper (e.g. Butcher and Rouse 2000, Schultz and Mwabu 1998).

Second, entrepreneurial skills may be absent in the population, as informal employment was

squashed under Apartheid (e.g. Kingdon and Knight 2004). Third, some unemployment

may be voluntary; a generous non-contributory pension program combined with the high

wages earned by the employed leave many unemployed individuals with networks capable

of supporting them (see Bertrand, Mullainathan, and Miller 2003 for labor supply eects;

Edmonds, Mammen, andMiller 2005 for network eects of pensions on living arrangements).

While it is clear that many adults are unemployed in South Africa, it is unclear what

adults are in fact doing. Labor force surveys in South Africa go to great lengths to measure

any economic activity, identifying as workers individuals who engage in unpaid household

work or tend household plots "even for only one hour" in the past week; this approach yields

the low employment numbers described above. A very natural response to this unemploy-

ment would be for many to create entreneurial work5. Yet row 2 of table 1 reveals that

only 6-8 percent of prime age South Africans are self-employed6. These numbers are tiny

compared to countries with similar levels of unemployment (e.g. Charmes 2000, Kingdon

4Kingdon and Knight (2006) advocate this broad unemployment measure in this context, as local wagesare more sensitive to that measure. O cial unemployment numbers include a broader range of ages andheld steady at 42% over the period of 2002-2004.

5This is particularly true as unemployment durations are very long, and there is some evidence thatsocial connections may be important to nd employment. Since jobs are scarce, job opportunities may beshared among very close relations (Magruder 2010, Seekings and Nattrass 2005), leaving individuals withpoor social connections with very limited opportunities to nd work.

6Among the majority black population, the corresponding gure is about 6% for both men and women

6

and Knight 2004). Moreover, what is perhaps most striking is that there are relatively few

employees of small rms in general. The remaining rows of table 1 reports the percent of

employees in each rm size category in South Africa. Particularly for men, we see very few

workers in rms of fewer than 5 employees. For comparison purposes, I also include similar

data from the 1995-96 Brazillian LSMS survey7. We see that, while unemployment is a

great deal higher in South Africa, the distribution of rm sizes looks fairly similar with

one big exception. What is missing in South Africa, compared to Brazil, are the small rms

with 2-4 employees.

Of the above explanations for high unemployment, one in particular which may suggest

minimal small-scale employment in a high unemployment context is well-enforced labor regu-

lation. The South African labor market is highly regulated, with a variety of legislated labor

standards as well as privately bargained arbitration decisions. Unlike many other developing

countries, South Africa is successful in enforcing labor and tax regulations on many small

rms; an inuential study found that the average business with fewer than 5 employees pays

nearly R14000 (about $2170) per employee in costs associated with tax and labor regula-

tions8(SBP 2005). Moreover, unions and rms can extend labor standard arbitration to all

workers in a given political district through bargaining councils. Small businesses, in par-

ticular, have advocated aggressively against the extension of these labor standards; in 2005

South African President Thabo Mbeki announced that small businesses would be granted a

blanket exemption from these bargaining council agreements within the year in his state of

the union address (Mbeki 2005). However, under pressure from trade unions and employers

organizations to the contrary, the government never enacted this blanket exemption (e.g.

Cosatu Rejects 2005). The fact both that the government would consider a legal change to

7It is not common for household surveys in developing countries to ask respondents about the size of therm they work for. Fortunately, the Brazillian LSMS is an exception. Brazil represents a particularly goodcomparison for South Africa as a country with a broadly similar income level and similarly extreme level ofinequality.

8This estimate is the average over complying and non-complying rms. The greatest contributor to thisestimated cost is VAT, though labor regulations are also important. Of course, small business respondentsto this survey may overstate compliance.

7

exempt small business and that it meant with strong opposition conrms the anecdotal and

survey evidence that these regulations are enforceable.

The potential of labor regulations to aect employment has been extensively explored

in economics and found generally mixed results; surveys of this literature are available in

Blau and Kahn (1999), Freeman (2009), and Nickell and Layard (1999). Much of the recent

literature (e.g. Bertrand and Kramarz 2002, Besley and Burgess 2004, Harrison and Scorce

2008) has adopted a dierence-in-dierences approach where a time series of data on the

legislative environment in states is summarized by a before and after period. Dierence in

employment trends between "treatment" states which adopt a policy and "control" states

which do not are then compared to estimate the eect of regulations on employment. A

second approach is to utilize a spatial discontinuity (e.g. Holmes 1998; Dube, Lester, and

Reich 2010), where neighboring counties or states are compared, under the assumption that

geographically proximate counties share similar labor markets and incentives to form labor

policy, but are dierentially exposed. Many existent labor regulation studies adopt some

elements of each of these approaches (e.g. Card and Krueger 1994), so that changes in trends

are compared across spatially proximate regions. The measure of each of these studies is

how comparable of a control group can be developed without causing small sample problems;

to determine which approach is best for South Africa will require a more careful description

of the labor regulations to be studied.

3 Industrial Bargaining in South Africa

Unions in South Africa can bargain with employers in two primary ways. The 1995 La-

bor Relations Act codies the right of employers to form employers organizations for their

particular industry and region and bargain with unions centrally; the labor standards which

result from this bargaining can then be applied to all employees working in the industry and

region which the bargaining council presides over. That is, if employer organizations and

8

unions decide to bargain centrally, then all employees who work within that geographical

region will work under the agreed-upon labor standards, regardless of their union status.

Unions and employers may also choose to bargain unilaterally, resulting in plant level agree-

ments (Bendix 2001). Both unilateral bargaining and centralized bargaining are observed

in a wide variety of industries and areas in South Africa, so that dierent industries in the

same location may be covered by dierent types of agreements, industries may be covered by

unilateral agreements in some locations and centralized agreements in others, and industries

in a particular location may be covered by centralized agreements in one year and not in

another.

It is encoded in law that bargaining councils must be representative of rms and employee

unions in their jurisdiction; however, the extent to which this law is enforced is unclear. The

o cial wording is that councils must be "su ciently" representative, leading to a great deal

of bureaucratic discretion and contention (primarily from small employers) as to whether

the agreements represent all interests (Bendix 2001). South Africas political structure is

that 354 magisterial districts are organized into one of 52 district councils; these in turn

comprise 9 provinces. In principle, there is not a strict criteria over which groupings of

magisterial districts can form a bargaining council; in practice, most bargaining councils

represent collections of magisterial districts which map to political boundaries, either na-

tional, provincial, or at the district council level. In the model outlined below, I follow the

empirical trend in presuming that other magisterial districts within the district council are

the natural bargaining partners in determining whether to form a bargaining council agree-

ment, while empirical analysis will standardize bargaining council units to eliminate any

potential endogeneity stemming from the choice of bargaining council size (and to determine

the "potential" bargaining council units for magisterial district-industry observations which

are not covered by a bargaining council).

Existing studies on the eects of arbitration on wages and unemployment in South

Africa have imperfect information on the presence of bargaining council agreements and

9

treat the endogeneity of union membership via industry and occupational xed eects, which

may be an imperfect control; these studies nd that unions receive very high wage premia,

particularly at the bottom of the income distribution (Schultz and Mwabu 1998), and that

bargaining councils exhibit a smaller, though still present, wage premium (Butcher and

Rouse 2001). However, since the right to bargain centrally is one which must be exercised

voluntarily, we may be concerned that bargaining council agreements exist systematically

in the industries, magisterial districts, and years in which local labor markets make them

particularly protable for the rms who pursue centralized bargaining.

Moll (1996) outlines a theoretical model discussing the implications of bargaining councils

for large and small rms. We may also imagine that large rm incentives depend on whether

the large rm is unionized. Suppose that, in the absence of a bargaining council agreement,

large unionized rms pay privately bargained wageswU, while large non-unionized rms

and small rms pay market wages (w). Under a bargaining council agreement, all would

pay the same bargaining council wagewBC

; following Moll (1996) in presuming that wU >

wBC > w, it is clear that operating costs decrease for large unionized rms and increase for

small rms and large non-unionized rms in the presence of a bargaining council agreement.

As the supply curves for the three types of rms shift, equilibrium changes. If small rms

have the lowest marginal products of labor (due to low capital stocks), we may imagine that

their supply curve shifts in by the largest margin, resulting in an increase in the residual

demand faced by the two types of large rms. Thus, large unionized rms benet from

less competition from small rms and lower wages, large non-unionized rms benet from

less competition from small rms but suer from higher wages, and small rms lose by the

greatest margins. The degree of these benets, and the degree to which small rms and large

non-unionized rms are punished by the bargaining council agreement, are functions of local

demand, local labor supply, production technologies at each rm size, and other local labor

market characteristics, as the changes in the demand faced by each type of rm will depend

on anything which inuences local supply and demand curves. While the intuition behind

10

these labor market responses is straightforward, I develop a model in the web appendix which

shows more formally that large unionized rms will increase employment in response to a

bargaining council agreement, while large non-unionized rms and small rms will decrease

employment.

The diering prot incentives that employers face, outlined above, are clear. Therefore,

the presence of a bargaining council agreement will clearly be related to some aggregation of

the private incentives of the large rms who initiate centralized bargaining. However, unions

could adopt a bargaining position which is more or less hostile to bargaining councils, so the

decision to pursue centralized bargaining may depend on both rm and union incentives.

Since most of South Africas unions are aggregated into three large nationwide alliances

who have centralized general policies towards bargaining councils (Bendix 2001), I model

the unions role in bargaining as a cost C of adopting the bargaining council agreement;

empirical analysis will be robust to any heterogeneity in this cost that is due to industry-

specic local labor markets or magisterial district characteristics9.

Suppose that, in the absence of a bargaining council agreement, large unionized rms in

magisterial district m earn prots Um, and that large non-unionized rms earn m: Further

suppose that all large rms each earn prots BCm in the presence of a bargaining council

agreement before paying cost C to the union, and that fraction m of the total Qm large rms

in magisterial district m are unionized. A bargaining council agreement is a collective result

of the preferences of large rms throughout a district council, thus, if magisterial district m

belongs to district council DC; bargaining council legislation is adopted if

Xm2DC

QmBCm C >

Xm2DC

Qmm

Um + (1 m) m

(1)

9Both COSATUs (South Africas largest union alliance) o cial positions on bargaining council agree-ments and the discussion of commentators (e.g. Bendix 2001) suggest that unions have some support forthese agreements due to the greater political support they receive from advocating for globally higher laborstandards. We may also imagine that unions have varying incentives related to local labor market hetero-geneity, for example, the amount of dues which can be received or local competition from uncovered workers.Empirical analysis will be robust to both of these possibilities.

11

Local labor demand, local labor supply, local production technologies, local unionization

rates, and local product demand all determine the result of this relationship. In places

where small rm production technologies are relatively ine cient, and large rms face little

competition, the incentives to form a bargaining council agreement are weakened, while in

places with a vibrant small rms sector, the incentives to enforce uniform wages may be

high. Any econometric investigation into the eect of bargaining councils on employment

and small rm employment would have to take these local labor market characteristics into

account.

4 Data and Descriptives

Data are drawn primarily from two sources. The South African Labour Force Surveys are

a nationally representative rotating panel conducted twice yearly from 2000 through the

present, each iteration surveying around 70,000 people. I use the September surveys from

2000-2003. Unfortunately, the panel aspect has not been well-maintained, with household

identiers not remaining consistent from wave to wave. As such, I aggregate data to the

magisterial district level and use it as a panel at that level. These data are not intended to

be representative at the magisterial district level and are not publicly released at that level

to prevent mistaken inference (on, for example, the extent of the variation in employment in

a particular magisterial district year to year). This concern, however, should not limit more

robust econometric analysis, so long as the degree to which the data are not representative is

unrelated to the variables of interest and local-level unobservable heterogeneity is properly

controlled for. While magisterial district identiers are not released, they can be inferred

from personal identication codes. These identiers remain unchanged since at least the

1997 October Household Survey, which published an association between number and local

municipality names10. From this list, I determine the magisterial district of each sampling

10Examining characteristics of magisterial districts between these two surveys reassures that the identiersare in fact unchanged. A change in coding in 2004 limits the sample to 2000-2003.

12

area, and determine the longitude and latitude for the population center of that magisterial

district. The unit of analysis in this paper will thus be the magisterial district; since

sampling weights are not designed to be representative at this level I do not use them.

Therefore, I measure employment in a given industry in a given magisterial district as the

number of people surveyed in that magisterial district who work in that industry11. We

may be concerned that very large magisterial districts have dierent labor markets from

their neighbors, and that we get little useful information out of small magisterial districts

where relatively few individuals were surveyed. I exclude the top and bottom two percent

of magisterial districts in terms of population from the analysis. Summary statistics of the

variables which will be used are included in table 2.

The presence of bargaining council agreements in a given year is revealed by the South

African Government Gazette, which publishes all agreements. A database compiled by the

author reveals which industries in which magisterial districts were covered by an agreement in

each year. This yields the outcome that 15 two-digit industries in South Africa are covered by

bargaining council agreements for at least some of the sample period. Of these, 7 industries

have cross-sectional variation in their coverage across the district councils of South Africa.

In 2003, 22% of prime-age African and Coloured workers in South Africa work in two-digit

industries where, in their magisterial district, some workers are covered by a bargaining

council agreement12. Dierent industries have dierent minimum eective scales, limiting

the potential for entrepreneurship in some industries. Table 3 reveals that 75% of the prime-

age African and Coloured self-employed in South Africa work in two-digit industries which at

11In a related point, it is not immediately obvious how to treat observations of 0 employment in somecategory in a particular town (of which there are many). On the one hand, these observations give usefuland important variation if bargaining councils are brutally eective, we may expect to see 0 small rmemployees in a particular town-industry. On the other hand, when I (ultimately) take log+1 as a measureof employment, the log operator strongly emphasizes observations which are 0. This concern is lessened bythe use of the simple count data rather than weighted counts the dierence between log(1) and log(301) isa lot more than the dierence between log(1) and log(2). Results which use the fraction of the populationemployed in that industry (available from the author) are similar in sign and in general more preciselyestimated than the logged results presented here.12The actual number of covered workers is probably lower, due to the aggregation at the 2 digit level.

Aggregation challenges are addressed below.

13

least sometimes have bargaining councils this suggests that these councils are being utilized

more in industries where small scale rms are economically viable. In contrast, only about

43% of workers overall are working in these industries. Looking within industries which at

least sometimes have a bargaining council, we see an even more interesting result. 48% of

employees who work in one of these industries are covered by a bargaining council agreement.

However, only 34% of self employed and 37% of small rm employees are covered, in contrast

with 69% of large rm employees that is, among industries which at least sometimes

have bargaining council agreements, places with bargaining council agreements have limited

small scale and self employment. The industries and the percentage of employment-weighted

magisterial districts covered in years 2000 and 2003 are listed in table 4. These bargaining

councils cover heterogeneous places in South Africa, and there is substantial variation, both

geographical and intertemporal. In the appendix, I present maps showing which magisterial

districts I code as always, sometimes, or never covered by a bargaining council agreement in

each industry, as well as a table which identies the number of magisterial districts which

add and remove bargaining councils in each industry in each year. Industries are quite

heterogeneous in their coverage patterns.

Industries are aggregated to the two-digit level. While many of the bargaining councils

are dened over two-digit industries, some are dened in a dierent way than the standard-

ized coding used in the labour force survey, and so only include subsets of those two-digit

industries (subsets which unfortunately do not always map to three-digit industries, the unit

reported in the surveys). This means that my measure of coverage includes individuals who

are actually working in distinct, uncovered jobs as well as covered employees. In principle,

this bias seems likely to result in conservative estimates due to measurement error, since the

bargaining council agreements only cover a fraction of the workers in the two-digit indus-

try. Two of the industries with variation end up in "other" categories. We might worry

that these categories are more heterogeneous than other two-digit designations, and that

the bargaining councils represented (hairdressing, laundry services, and contract cleaning)

14

represent a smaller fraction of the workers in the "other services" and "other business ac-

tivities" industries. Additionally, a third industry (electrical manufacturing) is very small

in scale (with only 25 small rm employees measured in South Africa across the 4 survey

years considered here), and covered almost everywhere. I exclude these three industries in

the analysis below, although similar analysis including these industries is available from the

author.

5 Econometric Model

The focus of this paper will be on estimating bargaining council eects on employment, rm

size, and wages. Model predictions suggested that bargaining councils may reduce overall

employment and small rm employment, but that they should have an ambiguous eect on

large rm employment. A linearized structural equation is given by

Yimt = + 1BCimt + Ximt + i + t + imt (2)

where Yimt may be employment, employment by rm size, or wages in industry i in

magisterial district m during year t; BCimt denotes the presence of a bargaining council

agreement, and Ximt are covariates including population and, in dierent specications,

magisterial district, magisterial district-year, or magisterial district-industry xed eects.

The discussion above suggests that the presence of a bargaining council agreement is related

to many characteristics of local labor markets, including labor supply, small rm production

technologies, etc. These are contained in imt; which may be correlated with BCimt and

other explanatory variables. Below, Ill consider several assumptions on imt to generate

a variety of potential control groups and illustrate the robustness of identied trends to a

variety of underlying assumptions.

Presuming we can adequately characterize imt and develop a comparable control group,

any analysis still requires a standardization on bargaining council size. As discussed above,

15

bargaining council agreements usually apply to all magisterial districts which belong to a

larger political entity, either a district council, province, or the entire nation. However,

in a few cases individual magisterial districts are added or subtracted from these groups in

the coverage of a bargaining council (usually either the biggest magisterial district or closest

neighbors of an adjoining district council). In fact, in a few cases bargaining councils cover

only a single magisterial district. Though these observations represent a small share of the

data, we may still worry about the implications of these observations for analysis, particu-

larly in the estimates which emphasize dierences across space at bargaining council borders

and which are the focus of this paper. Moreover, this fact forces consideration of poten-

tial bargaining council sizes for places with no current bargaining council in constructing

adequate control groups, which is necessary to identify the proper unit of analysis.

A direct solution to this problem is to treat bargaining council adoption as incomplete

take-up as discussed in the impact evaluation literature. In that literature, presuming we

have exogenous assignment of treatment but incomplete take-up among the treated, the

common solution is to instrument actual treatment status with assignment to treatment. I

follow that approach in this paper. Since most bargaining councils are assigned at larger

political boundaries than magisterial districts, I describe a magisterial district-industry-year

observation as eligible for the program if it belongs to a district council where at least

one magisterial district has a bargaining council agreement in that industry-year and use

that measure of eligibility as an instrument for program receipt. All rst stages are strong:

all t-statistics of bargaining council eligibility on bargaining council status are over 9.813. A

visual rst stage is presented in the appendix, which includes maps depicting both the actual

coverage and the instrumented coverage in each industry. The appendix also presents the

formal estimates of this rst stage in the papers main specications.

Second, regardless of the strategy used to solve the identication problem, two potential

sources of dependence among observations are well-known and relevant to this context. A

13Coe cients of bargaining council status on bargaining council eligibility range from 0.77 to 1.00.

16

rst challenge to evaluating programs which are implemented at aggregate levels is that

if individuals in a political district have correlated error terms, or there is autocorrelation

in the error, then OLS produces inconsistent standard errors (Bertrand et al 2004). The

standard solution is to cluster at the policy group level. Since bargaining councils vary on the

district council-industry level14 (there are 208 district council-industry groups and 52 district

councils in the estimation sample), this context avoids the small group number concerns

which have challenged some past studies of governmental policies (discussed in Donald and

Lang 2007). Secondly, the error term may be spatially autocorrelated (Conley 1999). As the

primary identication strategy will rest on the dierence between magisterial districts which

are physically proximate and those which are in the same political district, it is desirable

to construct standard errors which are robust to correlation amongst both groups. This

paper allows observations to be related if either they are close spatially or in either the

same district council in the primary analysis15. This is the more computationally intensive

procedure outlined in Cameron, Gelbach, and Miller (2006), and also a special case of the

Conley (1999) spatial errors if "economic distance" is dened as equal amongst individuals

who live either within a given physical distance or in the same District Council.

5.1 Econometric Benchmark 1: Dierence-in-Dierences

Our ability to estimate equation 2 depends on how we can characterize imt: Dierences-

in-dierences remain the dominant approach in the literature and make the assumption

that E [imtjXimt; i; y] = ~iDC ; allowing a simple District Council-Industry xed eect to14Several of the bargaining councils extend agreements to entire provinces, while others operate only on

the District Council level (and a small fraction operate for even smaller units). This makes it di cultto know, for certain, how to categorize observations (particularly for industries and towns which are notcovered by a bargaining council agreement). The results presented here presume that, since some DistrictCouncils unilaterally receive bargaining councils, this is the true observation level (implicitly, this assumesthat bargaining councils which exist react to considerations at the District Council level). An alternateassumption would be to assume that the observation unit is the province-industry level. Results whichmake this assumption and cluster simultaneously at the spatial-industry, province-industry, and town levelare similar and available from the author.15The less robust Spatial RD analysis clusters over space and within district council-industries as I describe

below.

17

eliminate the endogeneity concern. In that case, we can estimate

Yimt = + BCimt + Ximt + i + t + ~iDC + "imt (3)

where i, t; and ~iDC represent industry, year, and District Council-Industry xed ef-

fects, and Ximt includes a quartic in log population. Yimt variables include employment,

large rm employment, self employment, and small rm employment (small rms are dened

to be rms with fewer than 10 employees while large rms have more than 20), measured

as the log of that variable plus 1. Bargaining council status is instrumented with bar-

gaining council eligibility, as described in section 4.1. Panel A of table 5 performs the

dierence-in-dierences estimation on the main sample. All cells report the coe cient on

bargaining council presence for a given sample and dependent variable. Across the board, a

dierence-in-dierences suggests that eects of bargaining councils are negative and signi-

cant. As bargaining council agreements come into place, employment decreases by 7%, as

does small rm and self employment. However, we may be concerned that bargaining council

agreements are being adopted or eliminated endogenously in places with specic time trends

in employment or small rm employment, thus violating the necessary assumption for a

dierence-in-dierences estimation.

5.2 Econometric Benchmark 2: Spatial Regression Discontinuity

An alternate assumption notes that the endogenous characteristics of local labor markets

within a given industry are likely to be spatially continuous16, so long as migration and

trade are locally feasible. Formally, let R (m) denote the set of all magisterial districts

16It is possible that, as a legacy of Apartheid, some labor market characteristics may not be spatiallycontinuous due to poor infractrusctural connections (the Apartheid government did purposefully separateracial groups). If true, this would challenge identication in this paper. I address this issue in the section6.1 by running specications featuring either town-year or town-industry xed eects. Since barriers whichhave lingered since Apartheid should both eect all industries in a given town and all years within the sameindustry, these approaches will be robust to this concern. As it turns out, both yield consistent results,suggesting that lingering spatial barriers do not drive the empirical trends documented here.

18

within radius R of magisterial district m; Zimt be the vector [Ximt; BCimt; i; t] and ZiR(m)t

and niR(m)t denote the vectors of Zim0t and im0t; 8m0 2 R (m) : Then, spatial continuitysuggests that

EimtjZiR(m)t

= E

im0tjZiR(m)t

for R su ciently small and m0 2 R (m) : (4)

This assumption is similar to that made in standard regression discontinuity designs, sug-

gesting a spatial regression discontinuity. As with other regression discontinuity designs,

the appeal of spatial regression discontinuity is that we may expect (endogenous) local labor

market characteristics to vary smoothly over space. In this case, spatial continuity may be

confounded if rms are capable of resettling on one side of the policy regime, and spatial

discontinuity analysis will have to take this possibility seriously. Also as in other regres-

sion discontinuity designs, spatial RD will not estimate policy eects which vary smoothly

over policy borders; for example, if elevated wages aect equilibrium wage rates in adjacent

district councils, then the spatial RD will miss these eects.

Despite these concerns, there is a long history of economists exploiting spatial disconti-

nuities to identify the eects of regulations. Space is somewhat dierent from conventional

running variables employed in regression discontinuities in two ways. First, it is two dimen-

sional, and legal borders tend not to precisely collapse on any one of those two dimensions

(for example, they tend not to correspond to longitudinal lines). Second, there is less of an

a priori reason to suspect a systematic relationship between space and the outcome variable

than there is with many running variables. Most papers in this eld have resolved these

issues by collapsing the spatial data into a single dimension of "distance to the border",

following seminal work by Card and Krueger (1994) and Holmes (1998), and simply taking

the side of the border that an individual resides on as random for individuals within some

bandwidth of the border. Once we have reduced space to a single dimension, familliar

graphical and statistical analysis can proceed as in other regression discontinuity studies.

19

As noted above, population has an almost mechanical relationship with employment.

Thus, to implement the RD I rst estimate

Yimt = + g(popimt) + uimt (5)

where g () is a quartic in log population. By examining uimt as the unexplained variationwhich we may expect to be discontinous at bargaining council borders, we examine the

variation in employment which is not explained by dierences in survey population density

over the support of the running variable.

Heterogeneity in bargaining council size poses a second challenge. When we sum over

all bargaining councils, small bargaining councils are disproportionately represented among

observations which have bargaining councils and are close to the border, while large bar-

gaining councils are disproportionately represented among bargaining council observations

which are far from the border. If we want to examine the running variable over meaningful

lengths, we will be forced to interpret changes in employment over the support of distance

to the border as a combination of changes deriving from the eect of the bargaining council

and changes related to the composition of bargaining councils present at that dierence17.

To lessen the extent of these dierences, I exclude bargaining councils where no observations

are more than 50 miles from the border and which represent 17% of bargaining council ob-

servations, though similar gures are generated excluding bargaining councils which have no

observations more than 30, 75, or 100 miles from the border18.

Figure 1 (A), (B), (C), and (D) present a scatter plot of binned data for employment,

small rm employment, large rm employment, and self employment, where the X-axis

represents distance from the closest border of the industrys own bargaining council and

the scatter plot shows means for each 10-mile bin. Positive distances indicate that the

17I note that a similar trend does not occur at the non-bargaining council side of the border, as small andlarge bargaining councils both have observations which are near and far.18My data do not include a direct measure of bargaining council size. Here, I determine it as the maximum

distance any observation in that province-industry-year is from the border. The median bargaining councilhas observations up to 140 miles from the border.

20

observation is on the Bargaining Council side of the border. Each gure also overlays a tted

estimate of the trend with distance to the border, where the trend is a kernel-weighted local

polynomial regression estimated separately on each side of the border using an epanechnikov

kernel and a ten-mile bandwidth.

Figure 1 (A) reveals that for overall log employment, there is a clear drop in employment

at the bargaining council border, while gure 1 (B) shows a similar drop is even more

pronounced for small rm employment. Figure 1 (C) examines log large rm employment

and again nds evidence of a discontinuity at the bargaining council border, though it appears

to be driven by an increase in employment on the non-bargaining council side of the border

and is smaller in magnitude than the other discontinuities presented here. Finally, gure 1

(D) looks at self employment, and also nds evidence that self employment is reduced within

bargaining councils.

All of these RD estimates are larger in magnitude than the dierence-in-dierence esti-

mates. To assess statistical signicance, I tighten the bandwidth to 50 miles on either side

of the boundary, and estimate

Yimt = + BCimt + 1Distimt + 2BCimt Distimt + g (popimt) + "imt

where Distimt is the distance of magisterial district m from the bargaining council border

in industry i in year t, and other variables are as above. Panel B of Table 5 reports the

results of this exercise, where errors are clustered over space and at the Distict Council-

Industry level. Across the board, estimated eects from the Spatial RD design are large and

signicant, with Employment declining by 36%, Small Firm and Self Employment declining

by 30%, and Large Firm Employment decreasing by 23%. The standard errors are also

large, and we cannot rule out similarly sized eects to those in the dierence-in-dierence

estimation.

21

6 Spatial Fixed Eects

While the spatial discontinuity in this approach is both intuitive and transparent, it does

have two limitations. First, a border analysis not only compares individuals, magisterial

districts, or counties to proximate ones, it compares all magisterial districts on one side of the

boundary to all magisterial districts on the other (who may not be particularly proximate).

In other words, the guiding assumption to the above approach is that EimtjZiR(m)t

=

Eim0tjZiB(m)t

for B su ciently small and m0 2 B (m), where B (m) is the industrys

bargaining council border. Note that this assumption is somewhat dierent from assumption

4, as the geographic area spanned by B (m) may be quite large, even at small bandwidths,

as space has two dimensions. While using border-region xed eects can eliminate some

of this heterogeneity, it remains an imperfect approach as it introduces a discontinuity into

continuous space19.

Practically, this has three consequences. First, we reduce our precision substantially. In

contrast to a dierence-in-dierence approach or an approach where we compare only across

small regions of space, we fail to control for the very notable spatial and time-invariant

heterogeneity in employment. On top of that, restricting analysis to border regions removes

the ability of observations which are more distant from the border to identify the eects

of covariates with likely similar relationships across the full sample. If we are studying

employment, the most important of these may be population, as larger cities mechanically

employ more people. In practice, this loss of precision resulted in large standard errors

for the RD estimates. Second, we lose smaller bargaining councils from the data. There

are a number of ways to address this issue, but in general when we collapse space to a

single dimension and examine what happens as we move along that dimension, we will be

challenged by bargaining councils which are small. A number of these are important in South

19A dierent approach which solves this concern is presented in Dube, Lester, and Reich (2010). Thatstudy restricts the sample to border regions and uses contiguous county-pair xed eects, a similar dier-encing approach to that used below. However,as that study also restricts the sample to border counties, itloses the capacity of non-border regions to improve the precision of covariate estimates, as discussed below.

22

Africa, and the generalizability and interpretation of results would be aected by excluding

them altogether. Finally, the raw spatial RD ignores the panel nature of the data. When

bargaining council agreements change over time, a raw spatial RD will change the spatial

mapping to distance to the border, and lose track of the underlying labor market conditions

which may remain similar in a district over time as that district is moved around within

the mapping of "distance to the border." Given that dierence-in-dierences estimates

were more conservative than spatial-discontinuity estimates, it seems ideal to estimate a

specication which accounts for constant unobservable characteristics of magisterial districts.

This paper adopts the spatial xed eects (SFE) estimator introduced in Conley and

Udry (2008) and Goldstein and Udry (2008)20.The idea of this approach is identical to the

standard xed eects within estimator. For each observation, we can subtract o the mean

of observations which are spatially proximate.

Thus, if nR(m) represents the number of magisterial district-year observations in R (m) ;

we can represent the spatial xed eects estimator as

Yimt 1nR(m)

Xm02R(m);t0

Yim0t0 = 1

0@BCimt 1nR(m)

Xm02R(m);t0

BCim0t0

1A+ (6)

0@Ximt 1nR(m)

Xm02R(m);t0

Xim0t0

1A +t 1nR(m)

Xm02R(m);t0

t0 + imt 1nR(m)

Xm02R(m);t0

im0t0

If we assume that imt is spatially continuous and constant up to exible trend controls,

then Ehimt 1nR(m)

Pm02R(m);t0 im0t0 jBCiR(m)t; XiR(m)t; t

i= 0 and this within estimator

will consistently estimate 1 and : In this specication, identication is by spatial and

intertemporal discontinuity: outcomes are compared only against those of proximate neigh-

bors as are program status and covariates. This equation estimates whether, if a magisterial

districts bargaining council status is greater than its neighbors(i.e. the magisterial district

lies on the bargaining council-side of a border), then the magisterial district has less employ-

20In both of these papers, these spatial xed eects are used to control for unobserved soil quality variationwhich is presumed to be similar amongst nearby plots.

23

ment than its neighbors. Interior magisterial districts have a spatial deviation of zero in

bargaining council status, but still contribute to the estimation of employment eects of dif-

ferences in population and time trends. The analogous approach in conventional regression

discontinuity is to allow a exible relationship between the running variable and dependent

variable and examining a discontinuous jump at the eligibility cuto.

Finally, in addition to a benchmark spatial xed eects regression, I repeat all analysis

with magisterial district-year and magisterial district-industry xed eects in addition to

spatial xed eects. Magisterial district-year xed eects specications will be robust to any

dimensions in which the magisterial district is dierent from its spatially proximate neighbors

that year, which includes the possibility that the magisterial district may be politically

valuable to unions, the possibility that local neighbors are in fact distinct labor markets

due to terrain or infrastructural separations or legislation (presuming that all industries are

similarly aected by the long travel times or other disruptions to spatial continuity), or

the possibility of secular local time trends which aect all industries. Magisterial district-

industry xed eects control for any ways in which that industrys local labor market diers

from its spatially proximate neighbors which is constant over time21.

21Some care is required combining the spatial xed eects with other xed eects when other dimensionsof the panel are not balanced. In this paper, this happens when using the wage, tenure, and DC ratiosubsamples. In this case, the standard within estimator is biased, and so demeaning sequentially alongspatial and other dimensions is not consistent. Unfortunately, a simple adjustment such as that in Davis(2002) is not possible, as the spatial xed eects cannot be represented as a projection onto the columnspace of a number of dummy variables. As a result, I conduct these estimates using the full set of dummyvariables for the additional xed eects whenever the panel is unbalanced.The use of the full set of dummies adds an extra complication when combined with the space and political

jurisdiction clusters described below. As described in Cameron, Gelbach, and Miller (2006), clustering inmultiple dimensions can have the outcome that some diagonal elements of the estimated variance-covariancematrix are negative, which happens on the variance estimates associated with some of these nuisance para-meters in these subsamples. However, as inference is robust to clustering unilaterally on either dimension,and the standard errors on the coe cient of interest remain well-behaved (in the sense that relative magni-tudes between jointly clustering and clustering in either dimension stays similar), I continue to report these"correct" standard errors in tables here.

24

7 Spatial Fixed Eects Results

Column 1 of Table 6 reports the coe cients on the presence of a bargaining council agree-

ment on employment from several spatially dierenced estimations, where the estimation

equation is the instrumental variables analogue of equation 6, and spatial deviations in bar-

gaining council status are instrumented by spatial deviations in bargaining council eligibility.

In all equations, the spatial xed eect is taken at the 30-mile radius, so that each depen-

dent and independent variable represents deviations of variables between the observation of

interest and other observations in the same industry and within 30 miles, where distance is

determined by the great circle method. All estimations are conditional on a quartic in log

population and time xed eects, and all errors are clustered among observations across all

years of the same industry within 2 degrees of latitude or longitude, as well as among all

industries, magisterial districts, and years in the same district council. Having a bargaining

council agreement is associated with a signicant 8% reduction in log employment in the rst

row; including magisterial district-specic, magisterial district-year, or magisterial district-

industry xed eects keeps estimates of the bargaining council eect between 8-13%, and

always remains signicantly dierent from zero. These coe cients are quite stable despite

the very dierent identication assumptions: whether we look across industries at spatial

deviations in employment, or across time within industries, we draw very similar inferences

about the eects of bargaining councils22.

7.1 Firm Size Results

Here, I divide rms into four groups: large rms, with at least 20 employees; small rms, with

fewer than 10 employees; self-employment, and single-worker rms. Many self-employed in-

dividuals thus are also represented in the small rms and the single-worker rms categories.

From the model above, we expect bargaining councils to have the largest eects on employ-

22Very similar (and similarly precise) results are available if we use the fraction of the population employedin the industry instead of the logged specication.

25

ment in small rms. In principle, bargaining councils should have an ambiguous eect on

large rm employment, and the eect on self employment will depend on how many entre-

preneurs run larger small rms and the enforcement capacity of the bargaining council. If

most single-employee rms aspire to grow to multiple employees, or if single employees are

themselves paid a wage, it may be that bargaining council legislation reduces employment

in single-employee rms. However, since most single employee rms are owner-operated,

it seems likely that single-employee rms are primarily impacted through these dynamic

incentives, which may be weaker than the direct wage eects of the agreements. Therefore,

we may anticipate smaller eects among single-employee rms.

Columns 2 through 5 of table 6 reports the result of this analysis for each of these

dependent variables, where rows represent dierent xed eects specications (again, all

specications feature spatial xed eects in addition to the other noted xed eects). Here,

eects for small rms and self-employment are larger and consistently signicant. Con-

sistent with theory, bargaining councils reduce small rm employment substantially, with

bargaining council employment being associated with a 7-16% decline. This eect remains

very similar when we examine how spatial dierences vary within industries in a magisterial

district or magisterial district year and within an industry over time (just as in the overall

employment eects). Self employment similarly declines by 7-15%. Large rm employment,

in contrast, does not report a consistent eect. Coe cients are never signicant and are al-

ways smaller than small rm employment estimates. Similarly, single-employee employment

is not consistently related to bargaining council agreement status. This suggests that these

bargaining councils are most eective against small rms, but not single-employee rms, as

suggested by theory.

7.2 Wage Results

Of course, the stated purpose of the bargaining council legislation is to improve working con-

ditions rather then reduce employment. We can also ask if wages increase with bargaining

26

council agreements. This analysis uses the subsample with at least one wage observation,

which eliminates zero employment magisterial districts (and some with non-response to the

wage question). One consequence of the smaller sample is that the 30 mile radius, in con-

junction with various magisterial district-specic heterogeneity loses a lot of power; column

1 of table 7 indicates that we nd a 10-21% eect on wages at this radius, though standard

errors become large and the eect loses statistical signicance as we consider magisterial

district-year or magisterial district-industry xed eects. Column 2 repeats the analysis

with a wider 50 mile spatial radius for the spatial xed eects; at this larger radius the

magisterial district-year eects regain precision. Overall, industries represented by a bar-

gaining council in a magisterial district have 21% higher mean wages than the same industry

in neighboring magisterial districts, and 14% higher wages if we hold constant mean devi-

ations across industries in that year23. The motivation above suggested that small rms

should see larger wage increases than large rms, as large rms often must pay union wages

anyway. We can examine mean log wages for small rms (with fewer than 10 employees)

and large rms (with more than 20) separately, in columns 4 and 5. Consistent with the-

ory, wages in small rms are rising substantially, with (precisely measured) point estimates

around 12-20%. In contrast, large rm wages are if anything decreasing in response to

bargaining councils, consistent with the hypothesis that bargaining council wages are lower

than privately bargained ones (though errors are too large to reject a null hypothesis of a

zero eect). However, caution must be taken in interpreting wage estimates as a change in

wages for individual workers, because the composition of employees is changing. Column 5

controls for the fraction male, the average number of years of primary and secondary educa-

tion, a quadratic in average potential experience (age - education - six), and the fraction of

the workforce which is black, and nds that these controls attenuate wage eects by about

5 percentage points. In the appendix, I present estimated worker composition eects, and

23Since the wage data appear not to be su ciently dense for a 30 mile radius with town-year heterogeneity,I report the following wage regressions using the 50-mile spatial xed eects (30 mile radii give similar, butsometimes less precise, point estimates and are available from the author)

27

nd that the primary change in composition is that women and workers with low tenure are

being systematically disemployed by bargaining councils.

8 Robustness

There are two sets of concerns which could confound analysis. A rst set is econometric: the

spatial xed eects may be misspecied as they impose homogeneity assumptions over space

and over time. In the appendix, I weaken these two assumptions and nd that estimates are

robust to alternate spatial weights and to eliminating restrictions over time (e.g. estimating

spatial-industry-year xed eects). Second, there are two sets of economic endogeneity

explanations which are explored here. First, we may remain concerned that if a few labor

markets are very dierent from their neighbors and are important in determining bargaining

council status, then there may be discontinuities in labor markets which are systematically

related to bargaining council coverage. Second, similar statistical eects would be estimated

if rms simply resettle on the opposite side of a border or if employment is across-the-board

reduced by these bargaining council agreements, but these two regimes would have very

dierent policy implications.

8.1 Manipulation by Dominant Magisterial Districts

If some magisterial districts are very dierent from their neighbors and can determine bar-

gaining council policy, then we may not expect secular trends in these districts to be contin-

uous over space but they would be related to bargaining council policy. However, equation 1

makes clear that the presence of a bargaining council is due to some collaboration of magis-

terial districts in the same political district. If local labor markets are relatively continuous,

then nearby magisterial districts should have similar incentives to form a bargaining council

and the spatial xed eects approach solves the endogeneity problem. If they arent, then it

indicates that something about industry i in magisterial district m is dierent from industry

28

i in neighboring districts. If magisterial district m has much lower employment than other

magisterial districts in its District Council, then, as equation 1 makes explicit, magisterial

district m0s preferences should not be strongly reected in the presence or absence of a bar-

gaining agreement. In particular, if magisterial district m is discontinuously dierent from

its political neighbors in its incentive to form a bargaining council, then it will not be able

to enact its optimal choice. As such, our concern for endogeneity is minimized. However,

if a dominate share of industry i is located in magisterial district m; then this concern may

remain, and the presence of a bargaining council in magisterial district ms district council

may be a reection of these discontinuous labor market trends: In table 8, I repeat all esti-

mation with a sample of industries and magisterial districts where employment is no more

than 20% of employment in that industry in that district council on average.

Despite the smaller sample, precision increases and point estimates rise. Among magis-

terial district-industry groups which are too small to independently eect bargaining council

policy, we see employment fall by 12-16% relative to neighbors and other industries or other

years within the magisterial district. Consistent with the idea that endogeneity is minimized

in this subsample, results here line up precisely with theory, with the largest eects being on

small rms, smaller and marginally signicant eects on self employment, and consistently

small and insignicant estimates on large rm and single-employee rm employment. An

industry in a magisterial district which represents a small fraction of its countys employ-

ment can expect to see a 10-14% decline in small rm employment, a 5-9% decline in self

employment, and no change in its large rm or single-employee rm employment relative to

its neighbors.

8.2 Border Jumping

As with other regression-discontinuity estimators, manipulation of the running variable could

lead to mistaken inference. Here, we may be concerned that rms could relocate to a

magisterial district immediately on the opposing side of the border which would cause large

29

estimated spatial discontinuities but be a minimal concern for policy.

Some evidence against the importance of border-jumping was presented in Figure 1,

which indicated that employment levels remain depressed deep into the interior of bargain-

ing councils. A direct test (similar to Holmes 1998) would ask whether log employment

is dierent in a magisterial district if it is on the border of a bargaining council agreement

than otherwise. Specically, suppose we divide the bargaining council into several dis-

tance groups, so that we collect all magisterial district-industry-year observations which are

uncovered by a bargaining council but are either within 0-30 or 30-50 miles of a covered

magisterial district-industry-year observation, and similarly observations which are covered

by a bargaining council but within 0-30 or 30-50 miles of an uncovered magisterial district-

industry-year observation24. Then, we can test whether each of these border distance groups

are dierent from their counterparts in the same bargaining council regime by regressing

Yimt =Xk

1kDistGroupkimt + 2BCimt + t + i + "imt

However, local labor markets in border regions may systematically dier from those interior

regions. Two controls for spatial heterogeneity are used here. First, the magisterial district-

industry xed eects used earlier can still be used in this setting. These xed eects identify

control for any local labor market characteristics which remain constant over time and allows

us to examine simultaneously the eect of changing bargaining council status and changing

being on the border of a bargaining council. Second, I use xed eects at the District

Council-Industry-Year level. Here, border eects are identied o of magisterial districts

which are closer to the border than other magisterial districts in the same district council

(though bargaining council eects cannot be identied from this specication.)

24One could also measure distance to the border by geographic distance to the border (as opposed todistance to the nearest town on the other side of the border), as the measure in the Spatial RegressionDiscontinuity section is constructed. Estimates using that measure show less evidence of border jumpingand an identical estimate of the robustness of employment eects of bargaining councils. The measure ofdistance between two observations is preferred here as it more accurately reects the distance measure whichunderlies estimates.

30

Table 9 reports the results of this analysis, which reveals that border jumping is taking

place. When a bargaining council is formed near a given magisterial district but not including

that magisterial district, that magisterial district sees a large increase in employment (column

1). We similarly observe border jumping for large and small rms using only the spatial

variation, which reveal that having a bargaining council in your district council-industry-year

but being closer to the edge of the bargaining council regime is associated with some ight

of large rms (column 4, row 1), and that not having a bargaining council, but being near

magisterial districts that do, is associated with an increase in small rms (column 6, row 2).

In other words, we do nd evidence of manipulation - rms are purposefully locating outside

of the coverage of a bargaining council.

However, this fact is unrelated to the bargaining council eect documented in this paper,

as coe cients on bargaining council status are virtually unchanged by controlling for border

status (columns 1, 3, and 5) for overall employment, large rms, and small rms. This is in

part because border regions actually have more employment immediately on the bargaining

council side of the border as well as the non-bargaining council side (see the estimate on

column 1, row 1) , and in part because there are some fairly complicated spatial dynamics,

visible by comparing the eect of residing 30-50 miles out from the border to being within

30 miles of it. Regardless, we can conclude two things. First, border jumping is taking

place, suggesting that rms do prefer to resettle outside of the bargaining council regime and

oering supporting evidence that rms (and especially small rms) prefer to avoid bargaining

council agreements. Second, this eect is not inating our estimates of the employment

implications of bargaining councils.

9 Conclusions

Bargaining council agreements are the outcome of a complex bargaining process, which chal-

lenges inference as to their eects. This paper uses dierence-in-dierences, spatial regression

31

discontinuity, and spatial xed eects to demonstrate that bargaining councils are associ-

ated with about 8-13% lower employment in a particular industry, 10-21% higher wages,

and 7-16% less employment in small rms. Under spatial xed eects, these estimates are

further found to be robust to additional magisterial district, magisterial district-year and

magisterial district-industry xed eects. That is, an industry with a bargaining council

has about 8-13% less employment than its neighbors without a bargaining council. This is

true if we compare it to how dierent industries in the same magisterial district compare to

their neighbors, or if we compare how employment in that magisterial district and industry

changes over time with bargaining council status. Magisterial district-Industry observations

which employ a relatively small fraction of the employees in their District Council experience

the most severe consequences; that is, magisterial districts whose voices should receive little

weight in the decision to form a bargaining council are the most severely impacted by its

existence. The identication assumptions of spatial continuity can be weaker for these mag-

isterial districts if they dier substantially from their neighbors in their incentives to form

bargaining councils they will be unable to implement their desired bargaining council status

and so these estimates are particularly compelling. Moreover, while both small and large

rms appear also to prefer avoiding these restrictions, and hence resettle on the opposite

side of the border, this eect is unrelated to the estimated employment eect of bargaining

councils.

Eight to thirteen percent is a large decrease in employment in a given industry. By

means of comparison, Bertrand and Kramarz (2002) estimate that French entry regulations

reduce food retail employment by about 7%, Besley and Burgess (2004) estimate that labor

regulation reduced manufacturing employment in India by 7%, and Harrison and Scorce

(2008) nd that a 50% increase in the Indonesian minimum wage is associated with a 6%

employment reduction. The bottom end of the point estimates, then, is as large as these

eects of labor regulation found in other contexts. Similar centralized bargaining systems

exist in Western Europe and Latin America; these estimates suggest that if enforcement

32

is possible, and there is a large labor supply to low-productivity small rms jobs, then

these centralized bargaining structures may contribute strongly to unemployment. Balancing

this concern with improved labor conditions may be an important priority, particularly for

low and middle income countries. However, bargaining councils cannot explain all of the

unemployment problem in South Africa. 22% of employees work in two-digit industries in

places with bargaining council coverage, which corresponds to about 11% of the prime-age

population. If each of these industries were to increase employment by 8-13%, it would

cause a 0.88-1.43 percentage point total increase in employment. These eects are large and

should be of interest to policy makers. However, the South African unemployment situation

is severe enough that a one and a half percentage point increase in employment would leave

South Africa with a severe unemployment problem. So while the unemployment eects

of these policies are as big or bigger than other estimated labor regulation eects, other

problems still contribute to such high unemployment in South Africa. Spatially continuous

aspects of union behavior, labor market policies other than bargaining council agreements,

and the other voluntary and structural stories which may lead to high unemployment levels

may play an important role. The small rm eects is similarly large, and, unlike French

entry regulations (Bertrand and Kramarz 2002) hurts, rather than helps, small rms. This

policy is thus restricting small rm protability, in a context where the small rms sector was

already anemic. Once again, however, the small rms sector in South Africa is so minimal

that this 7-16% increase in these industries would leave small rm employment substantially

below global norms. Further research remains important to learn about the other potential

contributors to this problem.

References

(2005): Cosatu Rejects Plan to Bypass Centralized Bargaining,The Star, February 15, 6.

Banerjee, A., S. Galiani, J. Levinsohn, Z. McLaren, and I. Woolard (2008):Why Has Unemployment Risen in the New South Africa?,mimeo, MIT.

Bendix, S. (2001): Industrial Relations in South Africa. Juta and Co., Ltd, 4th edn.

33

Bertrand, M., E. Duflo, and S. Mullainathan (2004): How Much Should We TrustDierence-in-Dierences Estimates?,Quarterly Journal of Economics, 119(1), 249275.

Bertrand, M., and F. Kramarz (2002): Does Entry Regulation Hinder Job Creation?Evidence From the French Retail Industry,The Quarterly Journal of Economics, 117(4),13691413.

Bertrand, M., S. Mullainathan, and D. L. Miller (2003): Public Policy and Ex-tended Families: Evidence from Pensions in South Africa,World Bank Economic Review,17(1), 2750.

Besley, T., and R. Burgess (2004): Can Labor Regulation Hinder Economic Perfor-mance? Evidence from India,The Quarterly Journal of Economics, 119(1), 91134.

Blau, F. D., and L. M. Kahn (1999): Institutions and Laws in the Labor Market, inHandbook of Labor Economics, Volume 3, chap. 25, pp. 13991461. Elsevier-North Holland.

Butcher, K. F., and C. E. Rouse (2001): Wage Eects of Unions and Industrial Coun-cils in South Africa,World Bank Policy Research Working Paper No. 2520.

Cameron, A. C., J. B. Gelbach, and D. L. Miller (2006): Robust Inference withMulti-Way Clustering,mimeo, University of California, Davis.

Card, D., and A. B. Krueger (1994): MinimumWages and Employment: A Case Studyof the Fast-Food Industry in New Jersey and Pennsylvania, The American EconomicReview, 84(4), 772793.

Cardoso, A., and J. Gindin (2009): Industrial Relations and Collective Bargaining:Argentina, Brazil, and Mexico Compared, ILO Industrial and Employment RelationsDepartment Working Paper No. 5.

Carneiro, F. G. (1998): Productivity Eects in Brazilian Wage Determination,WorldDevelopment, 26(1), 139153.

Charmes, J. (2000): Informal Sector, Poverty and Gender: A Review of Empirical Ev-idence, http://www.wiego.org/papers/charmes3.pdf, Background Paper for the WorldDevelopment Report 2001.

Conley, T. G. (1999): GMM Estimation with Cross-Sectional Dependence,Journal ofEconometrics, 92(1), 145.

Conley, T. G., and C. R. Udry (2008): Learning About a New Technology: Pineapplein Ghana,The American Economic Review, forthcoming.

Davis, P. (2002): Estimating Multi-Way Error-Components Models with Unbalanced DataStructures,Journal of Econometrics, 106(1), 6795.

Donald, S. G., and K. Lang (2007): Inference with Dierence-in-Dierences and OtherPanel Data,The Review of Economics and Statistics, 89(2), 221223.

34

Dube, A., T. W. Lester, and M. Reich (2010): Minimum Wage Eects Across StateBorders: Estimates Using Contiguous Counties, Review of Economics and Statistics,92(4), 945964.

Edmonds, E. V., K. Mammen, and D. L. Miller (2004): Rearranging the Family?Income Support and Elderly Living Arrangements in a Low-Income Country,The Journalof Human Resources, 40(1), 186207.

Freeman, R. B. (2009): Labor Regulations, Unions, and Social Protection in Develop-ing Countries: Market Distortions or E cient Institutions?,NBER Working Paper No.14789.

Goldstein, M., and C. Udry (2008): The Prots of Power: Land Rights and AgriculturalInvestment in Ghana,The Journal of Political Economy, forthcoming.

Harrison, A., and J. Scorce (2008): Multinationals and Anti-Sweatshop Activism,The American Economic Review, forthcoming.

Holmes, T. J. (1998): The Eect of State Policy on the Location of Manufacturing:Evidence from State Borders,The Journal of Political Economy, 106(4), 667705.

Kingdon, G. G., and J. Knight (2004): Unemployment in South Africa: The Nature ofthe Beast,World Development, 32(3), 391408.

(2006): The Measure of Unemployment When Unemployment is High,LabourEconomics, 13(3), 291315.

(2008): Unemployment in South Africa, 1995-2003: Causes, Problems, and Poli-cies,Journal of African Economies, 16(5), 813848.

Liedholm, C., and D. Mead (1987): Small Scale Industries in Developing Countries:Empirical Evidence and Policy Implications,MSU International Development Paper No.9.

Liedholm, C., and D. C. Mead (1999): Small Enterprises and Economic Development.Routledge, London and New York.

Magruder, J. R. (2010): Intergenerational Networks, Unemployment, and PersistentInequality in South Africa,AEJ: Applied Economics, 2(1), 6285.

Mbeki, T. (2005): Address of the President of South Africa, Thabo Mbeki,at the Second Joint Sitting of the Third Democratic Parliament, Cape Town.,http://www.info.gov.za/speeches/2005/05021110501001.htm.

Moll, P. (1996): Compulsory Centralization of Collective Bargaining in South Africa,The American Economic Review, 86(2), 326329.

Nickell, S. (1997): Unemployment and Labor Market Rigidities: Europe versus NorthAmerica,Journal of Economic Perspectives, 11(3), 5574.

35

Nickell, S., and R. Layard (1999): Labor Market Institutions and Economic Perfor-mance,in Handbook of Labor Economics, Volume 3c, ed. by O. Ashenfelter, and D. Card,pp. 30293084. Elselvier-North Holland.

SBP (2005): Counting the Cost of Red Tape to Business in South Africa, Discussionpaper, Small Business Project.

Schultz, T. P., and G. Mwabu (1998): Labor Unions and the Distribution of Wagesand Employment in South Africa,Industrial and Labor Relations Review, 51(4), 680703.

Seekings, J., and N. Nattrass (2005): Class, Race, and Inequality in South Africa. YaleUniversity Press, New Haven and London.

Statistics South Africa (Producer) (1997): October Household Survey, 1997,South African Data Archive (distributor). www.nrf.ac.za/sada.

(2000-2003): South Africa Labour Force Surveys, September 2000, September2001, September 2002, September 2003, Pretoria, South Africa. South African DataArchive (distributor). www.nrf.ac.za/sada.

36

Table 1: Percentages of Employees in Firm Size Categories

South Africa BrazilMale Female Male Female

Percent Working 56.17 39.66 83.71 49.70Percent Self-Employed 7.87 5.98 27.49 24.31Number of Employees1 11.01 29.93 17.39 31.062-4 13.46 12.78 29.57 24.985-9 11.35 9.52 8.72 6.6310-19 14.06 12.64 7.08 5.5320-49 17.30 15.25 6.82 7.0250 + 32.82 19.88 30.43 24.78Notes1 South African Data are from the September 2003

labour force survey, while Brazillian data arefrom the 1995-95 LSMS survey.

2 Statistics are for 20-60 year old adults.

Table 2: Summary Statistics

mean sd min max NLog Employment 0.963 1.013 0.000 4.564 5048Log Large Firm Employment 0.330 0.599 0.000 3.664 5048Log Small Firm Employment 0.735 0.893 0.000 4.127 5048Log Self Employment 0.502 0.763 0.000 3.932 5048Log Single Firm Employment 0.346 0.651 0.000 3.829 5048Log Population 4.824 0.972 2.773 6.868 5048Bargaining Council 0.351 0.477 0.000 1.000 5048Mean log wage 5.216 0.787 1.569 9.798 2728Large Firm log wage 5.724 0.669 3.178 8.476 1260Small Firm log wage 5.025 0.832 1.569 9.798 2261Fraction Male 0.546 0.371 0.000 1.000 2728Potential Experience 23.957 9.245 0.000 66.000 2728Worker Education 8.495 2.704 0.000 12.000 2728Mean Log Tenure 1.194 0.774 0.000 3.761 2528

37

Table 3: Firm Size by Bargaining Council Status

Among All Industries Among BargainingCouncil Industries

Firm Size Never a BC BC Industry BC absent BC presentAll 56.67 43.33 51.8 48.2Self Employed 24.67 75.33 65.94 34.06Small Firms 54.53 45.47 62.65 37.35Large Firms 59.02 40.98 30.88 69.12Notes1 Summary statistics of adults aged 20-60 in September 20032 Bargaining Council Industries are covered by a Bargaining

Council Agreement at least sometimes

Table 4: Bargaining Council Coverage

Fraction Fraction of magisterialIndustry of workers districts covered

2000 2003Fishing 0.001 0 1Textile Manufacturing 0.020 1 1Metal Product Manf 0.015 1 1Electrical Machinery Manf 0.003 1 0.91Transport Equip Manf 0.005 1 1Furniture Manf 0.011 0.831 0.827Construction 0.101 0.343 0.204Retail Trade 0.256 0.156 0.217Hotels and Restaurants 0.058 0.336 0.266Land Transport 0.019 1 1Air Transport 0.001 1 0Other Business Acts 0.053 0.22 0.191Public Service 0.030 1 1Recreational/Cultural Act 0.005 1 1Other Service Activities 0.022 0.445 0.391Notes1 Fraction of workers is the average, 2000-20032 Fraction of magisterial districts is weighted by the working

population in that industry in each magisterial district

38

Table 5: Benchmark Estimates: Dierence-in-dierences and Spatial